Detecting Conceptual Abstraction in LLMs

Autor:	Regneri, Michaela, Abdelhalim, Alhassan, Laue, Sören
Rok vydání:	2024
Předmět:	Computer Science - Computation and Language Computer Science - Machine Learning
Druh dokumentu:	Working Paper
Popis:	We present a novel approach to detecting noun abstraction within a large language model (LLM). Starting from a psychologically motivated set of noun pairs in taxonomic relationships, we instantiate surface patterns indicating hypernymy and analyze the attention matrices produced by BERT. We compare the results to two sets of counterfactuals and show that we can detect hypernymy in the abstraction mechanism, which cannot solely be related to the distributional similarity of noun pairs. Our findings are a first step towards the explainability of conceptual abstraction in LLMs. Comment: Paper accepted at the LREC-COLING 2024 Conference (Paper ID: 1968) https://lrec-coling-2024.org/list-of-accepted-papers/
Databáze:	arXiv
Externí odkaz:	http://arxiv.org/abs/2404.15848 Zobrazit plný text záznamu View this record from Arxiv